Improved decomposition strategy based recurrent ensemble deep random vector functional link network for forecasting short-term electricity price

• Proposed a decomposition based recurrent ensemble deep random vector functional link neural network (REDRVFLN). • Successive variational mode decomposition (SVMD) decomposition performed for data decomposition. • The extracted SVMD modes are inputs to the REDRVFLN model. • Better than many other decomposition based approaches. • The ensemble approach with recurrent neurons improved the model accuracy. • Better performance offered for electricity price forecasting. In the present scenario, accurate short-term electricity price forecasting in a deregulated electrical market is highly essential. Therefore, this paper presents a novel approach using a recurrent ensemble deep random vector functional link neural network (REDRVFLN) with recurrent neurons hybridized with an improved decomposition strategy known as successive variational mode decomposition (SVMD) for short-term electricity price forecasting. This approach results in better generalization capacity, very simple structure and significant prediction accuracy. SVMD is used to transform the non-linear and non-stationary electricity price time series successively into a set of regular sub-series known as intrinsic mode functions (IMFs). Therefore this new approach does not need to know the number of modes apriori like the widely used variational mode decomposition (VMD), resulting with less computational complexity and more robustness in comparison to VMD. Further to handle temporal dependencies in the input sequence of the electricity price time series data a locally recurrent ensemble random vector functional link network (REDRVFLN) with stacked layers is used for processing the IMFs (features) as input sequence. REDRVFLN utilizes features from both the direct link and nonlinearly transformed features from preceding enhancement layers with locally recurrent neurons with feedback paths and fixed random weights. Further each layer produces an output by simple matrix inversion based on generalized least squares and all the outputs from different layers are combined by taking the median to obtain the final forecast thus producing a framework of both ensemble and recurrent deep learning simultaneously. The suitability of the REDRVFLN model is validated on two electricity markets (PJM, NSW) data that exhibit the lowest errors as compared to many single and decomposition based models. To assess the forecasting accuracy and predictive ability of the proposed approach, comparative forecasting performance of several benchmark and randomized models have been presented in this paper.